Several methods have been proposed for land covers classification of remote sensing images. However, for some complex and hard-to-access areas, collecting ground truth for supervised learning approaches is a hazard, expensive and time-consuming. Therefore, we focus on the automatic identification of land covers through specific features extracted from spectral bands. In this paper, we proposed a land covers identification method based on Dempster-Shafer theory, which is fully automatic to infer the semantic sense of labels without any manually labeling processing. Our contributions include an efficient method to extract vegetation through NDVI (Normalized Different Vegetation Index) and an automatic land cover identification using Dempster-Shafer theory.